[PATCH] adaptive lazy readahead

From: Suparna Bhattacharya <suparna@in.ibm.com>

From: Ram Pai <linuxram@us.ibm.com>

Pipelined readahead behaviour is suitable for sequential reads, but not for
large random reads (typical of database workloads), where lazy readahead
provides a big performance boost.

One option (suggested by Andrew Morton) would be to have the application
pass hints to turn off readahead by setting the readahead window to zero
using posix_fadvise64(POSIX_FADV_RANDOM), and to special-case that in
do_generic_mapping_read to completely bypass the readahead logic and
instead read in all the pages needed directly.

This was the idea I started with.  But then I thought, we can do a still
better job ?  How about adapting the readahead algorithm to lazy-read or
non-lazy-read based on the past i/o patterns ?

The overall idea is to keep track of average number of contiguous pages
accessed in a file.  If the average at any given time is above ra->pages
the pattern is sequential.  If not the pattern is random.  If pattern is
sequential do non-lazy-readahead( read as soon as the first page in the
active window is touched) else do lazy-readahead.

I have studied the behaviour of this patch using my user-level simulator.
It adapts pretty well.

Note from Suparna: This appears to bring streaming AIO read performance for
large (64KB) random AIO reads back to sane values (since the lazy readahead
backout in the mainline).

include/linux/fs.h

@@ -507,6 +507,8 @@ struct file_ra_state {
 	unsigned long prev_page;	/* Cache last read() position */
 	unsigned long ahead_start;	/* Ahead window */
 	unsigned long ahead_size;
+	unsigned long serial_cnt;	/* measure of sequentiality */
+	unsigned long average;		/* another measure of sequentiality */
 	unsigned long ra_pages;		/* Maximum readahead window */
 	unsigned long mmap_hit;		/* Cache hit stat for mmap accesses */
 	unsigned long mmap_miss;	/* Cache miss stat for mmap accesses */

mm/readahead.c

@@ -30,6 +30,7 @@ file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
 {
 	memset(ra, 0, sizeof(*ra));
 	ra->ra_pages = mapping->backing_dev_info->ra_pages;
+	ra->average = ra->ra_pages / 2;
 }
 
 EXPORT_SYMBOL(file_ra_state_init);
@@ -380,9 +381,18 @@ page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
 		 */
 		first_access=1;
 		ra->next_size = max / 2;
+		ra->prev_page = offset;
+		ra->serial_cnt++;
 		goto do_io;
 	}
 
+	if (offset == ra->prev_page + 1) {
+		if (ra->serial_cnt <= (max * 2))
+			ra->serial_cnt++;
+	} else {
+		ra->average = (ra->average + ra->serial_cnt) / 2;
+		ra->serial_cnt = 1;
+	}
 	preoffset = ra->prev_page;
 	ra->prev_page = offset;
 
@@ -449,8 +459,12 @@ page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
 			  * accessed in the current window, there
 			  * is a high probability that around 'n' pages
 			  * shall be used in the next current window.
+			  *
+			  * To minimize lazy-readahead triggered
+			  * in the next current window, read in
+			  * an extra page.
 			  */
-			ra->next_size = preoffset - ra->start + 1;
+			ra->next_size = preoffset - ra->start + 2;
 		}
 		ra->start = offset;
 		ra->size = ra->next_size;
@@ -468,11 +482,27 @@ page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
 		}
 	} else {
 		/*
-		 * This read request is within the current window.  It is time
-		 * to submit I/O for the ahead window while the application is
-		 * crunching through the current window.
+		 * This read request is within the current window.  It may be
+		 * time to submit I/O for the ahead window while the
+		 * application is about to step into the ahead window.
 		 */
 		if (ra->ahead_start == 0) {
+			/*
+			 * if the average io-size is less than maximum
+			 * readahead size of the file the io pattern is
+			 * sequential. Hence  bring in the readahead window
+			 * immediately.
+			 * Else the i/o pattern is random. Bring
+			 * in the readahead window only if the last page of
+			 * the current window is accessed (lazy readahead).
+			 */
+			unsigned long average = ra->average;
+
+			if (ra->serial_cnt > average)
+				average = (ra->serial_cnt + ra->average) / 2;
+
+			if ((average >= max) || (offset == (ra->start +
+							ra->size - 1))) {
 				ra->ahead_start = ra->start + ra->size;
 				ra->ahead_size = ra->next_size;
 				actual = do_page_cache_readahead(mapping, filp,
@@ -481,6 +511,7 @@ page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
 						actual, orig_next_size);
 			}
 		}
+	}
 out:
 	return;
 }